Microsoft Launches Paza To Advance Speech Recognition For Low-Resource Languages

Microsoft Research has launched Paza, a new initiative aimed at advancing automatic speech recognition (ASR) for low-resource languages, starting with African languages that have historically been underserved by mainstream AI systems.

As voice technology becomes central to accessing information and services, many languages remain poorly supported, limiting participation in the digital economy for communities that rely primarily on spoken communication. Paza addresses this gap by combining a transparent benchmarking platform with speech models built and tested for real-world use.

“Language should not be a barrier to access in the digital age,” said Mercy Muchai, Research Engineer II and one of the project leads. “With Paza, we are co-creating speech technology with the communities who use it, ensuring that the voices of speakers of low-resource languages are heard, understood, and meaningfully included.”

At the heart of the initiative is PazaBench, the first evaluation platform dedicated to low-resource languages. The platform covers 39 African languages and evaluates 52 state-of-the-art speech and language models, including the newly released Paza ASR models for six Kenyan languages. By providing a standardized and publicly accessible framework, PazaBench allows researchers and developers to identify performance gaps and accelerate improvements in speech technology for underserved communities.

Alongside the benchmark, Microsoft Research is introducing fine-tuned ASR models optimized for Swahili and several Kenyan languages, including Dholuo, Kalenjin, Kikuyu, Maasai, and Somali. Developed through strong community engagement, the models were tested in everyday environments including low-bandwidth and noisy conditions to ensure reliability and usability beyond the lab.

The initiative builds on previous collaborations supporting underserved communities through AI-driven tools, emphasizing the importance of systems that account for local accents, dialect variations, connectivity challenges, and cultural context. Paza, a Swahili word meaning “to project” or “to raise your voice,” reflects this commitment to ensuring speech technology works for everyone.

“Real-world speech recognition must work for everyone, not just a few widely spoken languages,” said Kevin Chege, Machine Learning Engineer on the project. “By testing models with community members  including farmers using everyday mobile devices, we’re building systems that reflect real usage conditions and deliver practical value.”

Looking ahead, Microsoft Research plans to expand PazaBench to include additional low-resource languages and publish practical guidance for responsible dataset development, model fine-tuning, and real-world evaluation. By openly sharing benchmarks and methodologies, the initiative aims to foster collaboration across academia, industry, and local communities, accelerating more inclusive AI innovation.